What the documents look like
Cross-border payment risk work centres on card-network disputes and receiving accounts. The material set is dominated by chargeback notices, fraudulent-order lists, invoices and proof of delivery, which differs from parcel-status fulfilment records. For the inventory procurement and supplier records task, supplier names, purchase batch numbers, invoice numbers, received quantities and signed delivery dates must sit inside one retrieval scope so that a risk specialist can check how an order relates to a frozen amount. Records update daily and per order, so retrieval frequently faces several documents for the same supplier across different batches, with similar wording and different field values. The knowledge base search node can associate one or more knowledge bases that use the same vector model; knowledge bases with different vector models cannot be associated together.
Constraints at the retrieval step
Retrieval only returns candidate content for a question. It does not perform authorization checks. Cited content is returned as an array, and that array can have a length of zero, so the output path still completes when nothing is found. The minimum relevance value is a filter threshold, not a ranking rule, and it does not isolate permissions or business entities. When question optimization, hybrid retrieval, image retrieval or result reranking is active, the final result is merged from several paths and need not follow a single vector similarity order. Hybrid retrieval merges semantic and full-text results through RRF, reranking runs when configured, and image retrieval depends on the relevant image input and model capability. These paths do not mean every request executes all of them.
Hypothetical paper test case: the source material reads "Supplier A issued invoice INV-0310 on 10 March 2025, purchase batch PO-2025-018, quantity 1,200 units, corresponding delivery receipt SIGN-0312, signed date 12 March 2025." The draft sentence to be judged is "Supplier A's PO-2025-018 has completed signed delivery and settled the payment." The check is limited to this supplier, this batch and this invoice field. The source states only the signed date. It does not state the payment settlement status, so the second half of the draft sentence is unsupported.
Configuration and implementation checks
| Configuration item | Hypothetical starting point for this trial | Reason and comparison method |
|---|---|---|
| Associated knowledge bases | Associate only the inventory procurement and supplier records base | Check each base's vector model identifier first, exclude those that differ, and do not run cross-model retrieval |
| Search mode | Hybrid retrieval | Invoice numbers and batch numbers are exact queries while supplier names are semantic queries; compare the two recall paths |
| Citation limit | Measured in tokens, starting at a smaller value that can hold one complete invoice record | Test with long and short invoices and delivery receipts for content completeness, then adjust |
| Minimum relevance | Label relevant and irrelevant samples first, then set the threshold | This value applies only in semantic retrieval or with result reranking, and does not act as a permission filter |
| Question optimization | Enable for multi-turn follow-up scenarios | Compare the original question with the optimized question and check whether the supplier and batch subject were altered |
| Result reranking | Enable as configured | Rerank scores are used for filtering; confirm the model is available |
Three common errors
- Treating a retrieval hit as an authorization conclusion. Error: retrieving invoice INV-0310 is taken to mean the batch payment is settled. Check: input the draft sentence and the source material, compare the signed date field with the settlement status field, observe that the source does not state settlement, and record the result as unsupported.
- Treating minimum relevance as entity isolation. Error: raising the threshold is assumed to exclude other suppliers' records. Check: input a query containing two supplier names, compare the returned content with the target supplier scope, observe whether unrelated records are still recalled, and note that the threshold does not carry that function.
- Associating knowledge bases with different vector models. Error: bases using different vector models are associated together. Check: inspect each base's vector model identifier, exclude those that do not match, and record the actual observation without inferring runtime behaviour.
How to verify the setup
- Input the hypothetical source material and the draft sentence, compare the signed date field with the payment settlement field, observe that only the former appears in the source, and mark the second half of the draft sentence as unsupported. The judgment holds under that condition.
- Input a paper-level empty result array and check whether the downstream design handles zero citations. Do not claim that retrieval under a fictional rule necessarily returns an empty array.
- Input two query types, invoice number INV-0310 and a supplier name, compare the semantic and full-text recall content, observe whether the number query hits the corresponding record and whether the semantic query recalls nearby batches, and in the failure branch record the miss and keep the evidence gap open.